Detecting Diverse Seizure Types with Wrist-Worn Wearable Devices: A Comparison of Machine Learning Approaches

Area of Science:

• Biomedical Engineering
• Neurology
• Machine Learning

Background:

• Current seizure detection methods often focus on generalized tonic-clonic (GTC) seizures.
• There is a need for wearable technology to detect a wider range of seizure types, including focal and subclinical seizures.
• Machine learning (ML) offers potential for analyzing complex biosignal data from wearable devices.

Purpose of the Study:

• To assess the feasibility of using wrist-worn wearable devices with ML for detecting diverse seizure types.
• To evaluate the effectiveness of different ML models and data processing strategies for seizure detection.
• To identify limitations and areas for improvement in wearable-based seizure detection systems.

Main Methods:

• Twenty-eight patients were monitored using Empatica E4 wrist-worn devices during inpatient video-EEG.
• Devices collected accelerometry, blood volume pulse, electrodermal activity, skin temperature, and heart rate data.
• XGBoost, deep learning (LSTM, CNN, Transformer), and ROCKET models were evaluated using leave-one-patient-out cross-validation with varying segment lengths and feature sets.

Main Results:

• Generalized tonic-clonic (GTC) seizures were detected most reliably (AUROC = 0.86, SW-Recall = 0.81).
• Hyperkinetic and tonic seizures had high recall but also high false alarm rates.
• Subclinical and aware-dyscognitive seizures showed the lowest recall and highest false alarms; longer segments (60s) reduced false alarms.

Conclusions:

• Wrist-worn wearables combined with ML can detect seizures beyond GTC types, but effectiveness for non-motor seizures is limited.
• Model selection, feature sets, and segment length optimization are crucial for clinical utility.
• Minimizing false alarms is essential for real-world adoption of wearable seizure detection technology.